EEG-based Brain-Computer Interfaces (BCIs): A Survey of Recent Studies on Signal Sensing Technologies and Computational Intelligence Approaches and their Applications
(2001.11337v1)
Published 28 Jan 2020 in eess.SP, cs.AI, and cs.HC
Abstract: Brain-Computer Interface (BCI) is a powerful communication tool between users and systems, which enhances the capability of the human brain in communicating and interacting with the environment directly. Advances in neuroscience and computer science in the past decades have led to exciting developments in BCI, thereby making BCI a top interdisciplinary research area in computational neuroscience and intelligence. Recent technological advances such as wearable sensing devices, real-time data streaming, machine learning, and deep learning approaches have increased interest in electroencephalographic (EEG) based BCI for translational and healthcare applications. Many people benefit from EEG-based BCIs, which facilitate continuous monitoring of fluctuations in cognitive states under monotonous tasks in the workplace or at home. In this study, we survey the recent literature of EEG signal sensing technologies and computational intelligence approaches in BCI applications, compensated for the gaps in the systematic summary of the past five years (2015-2019). In specific, we first review the current status of BCI and its significant obstacles. Then, we present advanced signal sensing and enhancement technologies to collect and clean EEG signals, respectively. Furthermore, we demonstrate state-of-art computational intelligence techniques, including interpretable fuzzy models, transfer learning, deep learning, and combinations, to monitor, maintain, or track human cognitive states and operating performance in prevalent applications. Finally, we deliver a couple of innovative BCI-inspired healthcare applications and discuss some future research directions in EEG-based BCIs.
This survey paper, "EEG-based Brain-Computer Interfaces (BCIs): A Survey of Recent Studies on Signal Sensing Technologies and Computational Intelligence Approaches and their Applications" (Gu et al., 2020), provides a comprehensive review of advancements in EEG-based BCIs between 2015 and 2019. It focuses on signal sensing technologies, computational intelligence approaches (including machine learning, fuzzy models, and deep learning), and their real-world applications, particularly in healthcare.
The paper begins by introducing Brain-Computer Interfaces (BCIs) as direct communication pathways between the human brain and external devices. It categorizes BCIs into active/reactive and passive types and outlines their diverse application areas, such as human-computer interaction (HCI), gaming, neurocomputing for pattern recognition, and healthcare (e.g., assistive devices, disease detection). It then discusses various brain imaging techniques, highlighting the advantages of non-invasive Electroencephalography (EEG) – portability, high temporal resolution, and lower cost – despite its lower signal-to-noise ratio (SNR) and spatial resolution. Key EEG phenomena like Event-Related Potentials (ERPs, e.g., P300), Rapid Serial Visual Presentation (RSVP), Steady-State Visual Evoked Potentials (SSVEP), and Psychomotor Vigilance Task (PVT) are also introduced.
The authors state their contribution as a systematic summary of recent (2015-2019) BCI studies, covering EEG sensing technologies, signal enhancement, machine learning algorithms (including interpretable fuzzy models and transfer learning), deep learning methods, and BCI-inspired healthcare applications, aiming to fill gaps left by previous surveys.
Key areas covered in the survey include:
1. Advances in Sensing Technologies
The survey reviews progress in EEG sensor technology, contrasting traditional wet electrodes (requiring conductive gel) with modern dry electrodes, which enhance usability and comfort for real-world applications. It highlights research on dry sensors with signal quality comparable to wet electrodes. A comprehensive table and figure showcase various commercially available EEG devices, detailing attributes like wearability, sensor type (wet/dry), number of channels, sampling rate, and transmission type. Devices range from low-resolution (1-32 channels) for frontal/temporal coverage to high-resolution (>128 channels) for comprehensive scalp coverage.
2. Signal Enhancement and Online Processing
Effective BCI implementation relies on clean EEG signals. The paper discusses:
Artefact Handling: Techniques to remove noise from eye blinks, eye movements, and muscle activity. Blind Source Separation (BSS) algorithms like Principal Component Analysis (PCA), Canonical Correlation Analysis (CCA), and particularly Independent Component Analysis (ICA) are emphasized. Artefact Subspace Reconstruction (ASR) and combinations like EEMD-CCA for muscle artefact removal are also mentioned.
Toolboxes: EEGLAB and its extensions (e.g., AAR, MARA, clean_rawdata, ADJUST) are highlighted as widely used tools for EEG data processing and artefact removal.
Online Processing: The challenges and advancements in real-time EEG data collection, preprocessing, artefact rejection, and analysis. Systems like the Source Information Flow Toolbox (SIFT) and Real-time EEG Source-mapping Toolbox (REST) enabling real-time cognitive state classification are discussed.
3. Machine Learning and Fuzzy Models in BCI Applications
The survey covers various computational intelligence approaches:
Machine Learning Overview: It describes supervised and unsupervised learning, and common models in BCI such as linear classifiers (LDA, SVM), neural networks (MLP), non-linear Bayesian classifiers (HMM), nearest neighbor classifiers (kNN), and classifier combinations. The typical pipeline of EEG data preprocessing, feature extraction, and machine learning is illustrated.
Transfer Learning (TL): Addresses the issue of inter-subject and intra-subject variability in EEG data.
Need for TL: Reduces calibration time and improves generalization across different subjects, sessions, or tasks.
Types of TL: Inductive, transductive (including domain adaptation and covariate shifting), and unsupervised TL.
Applications in BCI: Transferring information task-to-task, subject-to-subject, session-to-session, and even headset-to-headset (e.g., using LST, wAR, AwAR, FWET, label alignment).
Interpretable Fuzzy Models: To address the "black box" nature of many machine learning models.
Concepts: Fuzzy sets, Fuzzy Inference Systems (FIS) for extracting "If-Then" rules, fuzzy integrals for data fusion, and hybrid models like Fuzzy Neural Networks (FNNs, e.g., SONFIN).
EEG-based Applications: Extending CSP filters with fuzzy sets for regression, using fuzzy membership for entropy calculation, integrating fuzzy logic with domain adaptation (OwARR), and fuzzy fusion for motor imagery BCI.
4. Deep Learning Algorithms with BCI Applications
Deep learning models that jointly learn features and classifiers directly from data are extensively reviewed:
Convolutional Neural Networks (CNNs): Their architecture (convolutional, pooling, fully connected layers) is well-suited for capturing spatial dependencies in EEG. Applications are detailed in:
Motor Imagery (MI) classification (often combined with CSP).
Emotion recognition (using raw EEG or extracted features like PSD, PLV, PCC, PLI; DGCNN).
Generative Adversarial Networks (GANs): Used primarily for EEG data augmentation to address insufficient training data. WGAN-GP and CWGAN are mentioned for synthesizing EEG data to improve classification in tasks like emotion recognition. GANs are also explored for EEG super-resolution.
Recurrent Neural Networks (RNNs) and Long Short-Term Memory (LSTM): Suitable for modeling temporal dependencies in EEG time-series.
Applications: SSVEP classification, auditory stimulus classification, visual object classification, individual identification, hand motion identification, sleep staging (e.g., SleepNet), and emotion recognition. Often combined with CNNs (e.g., DeepSleepNet, hierarchical LSTM for decision prediction).
Deep Transfer Learning (DTL): Combines deep learning with transfer learning to leverage pre-trained models or features.
Categories: Instance-based, mapping-based, network-based, and adversarial-based DTL.
Applications: Predominantly in MI-EEG classification (transferring knowledge subject-to-subject or session-to-session using pre-trained CNNs like VGG-16), attention detection, and imagined vowel pronunciation. GANs are also used in DTL for domain adaptation (e.g., DupGAN).
Adversarial Attacks: The vulnerability of deep learning models in BCI to deliberately crafted, small perturbations that can mislead the model. The paper discusses white-box, black-box, and gray-box attack scenarios and research on attacking popular CNN models for EEG (EEGNet, DeepCNN, ShallowCNN) and developing more practical attack methods (e.g., active learning for black-box attacks, universal adversarial perturbations).
5. BCI-based Healthcare Systems
A significant portion of the survey is dedicated to the application of EEG-based BCIs in healthcare:
Neurodegenerative and Brain Disorders:
Epilepsy: Automated seizure detection and prediction using RNNs (GRU, LSTM) and CNN-LSTM algorithms.
Parkinson’s Disease (PD): CNN-based diagnosis and Echo State Networks (ESNs) for classifying EEG from REM Sleep Behavior Disorder (RBD) patients (a risk factor for PD).
Alzheimer’s Disease (AD): CNNs for classifying mild cognitive impairment and AD from EEG; machine learning with multiple EEG biomarkers to enhance AD classification.
Schizophrenia: Machine learning with sensor/source level EEG features and deep learning architectures for diagnosis.
Motor Impairment Rehabilitation: Non-invasive EEG-based BCIs to aid hand movement and support post-stroke motor rehabilitation by rewarding cortical action.
Other Healthcare Areas:
Migraine: ANN-based classification and fuzzy entropy for studying migraine phases (pre-ictal, inter-ictal).
Pain: Investigating correlations between EEG spectral patterns and chronic pain; BCI for phantom limb pain control.
Depressive Disorders: CNNs and DTL for depression recognition from EEG, identifying critical spectral information and useful electrode locations (e.g., temporal areas).
Discussion and Conclusion
The paper concludes by acknowledging the significant progress in EEG-based BCIs due to advances in sensor technology, signal processing, and computational intelligence. However, challenges in real-world usability, prediction/classification capability, and stability in complex scenarios remain.
Future research directions highlighted include:
Further development of cost-effective, comfortable, and high-quality dry EEG sensors.
Integrating BCI with other physiological signals (hybrid BCIs) to improve accuracy.
Combining Augmented Reality (AR) with EEG-BCIs for immersive applications (e.g., using SSVEPs with AR glasses).
Developing adaptive BCI training leveraging TL and DTL.
Creating robust defenses against adversarial attacks on BCI systems.
Exploring closed-loop BCI methods with reinforcement learning for therapy and improved training.
Overall, the survey provides a detailed overview of the state-of-the-art in EEG-based BCIs from 2015-2019, emphasizing practical implementation aspects, diverse applications, and future challenges in this rapidly evolving interdisciplinary field.